AI eats Software
When do we reach AGI?
The quest for AGI is a journey shrouded in anticipation – we’re all waiting for a signal, perhaps from Sam Altman, to declare it’s here. I’ve begun to realize that we’ve gotten too lost in the minutiae of what it means to seriously consider the fact that it might already be here. I mainly subscribe to Sam’s definition that we’ve reached AGI when machines outperform the median human in any task. But why wait for the “General” condition in Artificial General Intelligence to be satisfied? There’s already many domains where LLMs are better than the median human, such as coding, which is why I think it’s time we declare coding-level AGI and seriously consider what that means for the future of technology companies.
LLMs will create full-stack factories
If you’re not convinced that we’ve already achieved coding AGI, this is for you. Initially, I pigeonholed ChatGPT as a tool for simple Python tasks. But by mid-2023, my trust evolved, and I began harnessing it to craft complex full-stack apps. Despite my limited iOS expertise, ChatGPT enabled me to rebuild an entire app that I previously paid for, requiring less than ten lines of my own code for minor tweaks. Being an engineer definitely gives me an advantage in prompting, but nothing I did was rocket science. I described my needs, asked what the data model should look like, kickstarted the project in XCode, and detailed the UI/UX for each segment. When the output wasn't perfect, I just asked differently but all in plain English and all in the way you might ask any engineer to do it if you knew they would not punch you in the face for asking too much.
My “Aha!” moment here is the realization that we’re on the brink of turning language models into full-stack app factories, even without further advancements. The missing piece is an orchestration layer that seamlessly integrates LLM outputs, manages files, and handles deployment. This isn't just a tractable problem—it's inevitable that many companies will work on this given skyrocketing software salaries.
Why isn’t this obvious?
This should be obvious to every technical person, right? Except, it’s not. The more I started telling people I was able to do this the more I realized it was not obvious that this is even possible today. The main reasons people seem unaware:
They’re non-technical and are not building with ChatGPT
They’re an engineer at a big company and the most they’ve used is some copilot tool
They haven’t learned how to prompt
Most engineers I’ve talked to in group 2 are still in the stage of believing ChatGPT is good for one-off python functions. These same engineers tend to believe that it should be horrible at architecture design because that’s the “important stuff.” I find this funny because architecture is less of an art and more of a science where there are good and bad patterns. If you’ve ever given an architecture interview for a tech company, you’d know that there’s a finite set of acceptable answers. Yes, there are software systems that get complex enough that they require innovation but this is not true for most SaaS products.
The Future of SaaS
In the next 2-3 years, we should expect the existence of some product that generates any application for you. Let’s call it Alfred. It’s tempting to believe that Alfred will just generate simple web apps but it’s more likely that market forces will conspire to make Alfred incredibly robust – there’s just too much at stake. What does the world look like when Alfred exists? I believe that the most direct implication is that every SaaS category will be a red ocean. The reality today is that the amount of SaaS competition that exists is limited by the amount of development resources. Some people might argue that outsourcing has brought down the costs of software development already (so Alfred pretty much exists) but this notion is flawed. Outsourcing costs a lot of money and if you’ve ever outsourced you’d realize it’s hit or miss. ChatGPT doesn’t really code at the level of an outsourced engineer, it codes at the level of some of the best Silicon Valley engineers, which means that what was once a very limited reagent will be in infinite supply.
For some categories of Vertical SaaS, it may no longer make sense to expect venture scale outcomes. If your product is not working, no one cares. If people gain a whiff that your product is working, there will be thousands of people motivated to ask Alfred to build them a version of it.
Does the tech even matter?
Maybe you don’t believe the tech is even the differentiator. In your mind, the differentiator has always been the sales process, the brand, or some other non-technical dimension. I think this can still be true and that winning products might be the ones with the best brand, but I believe what’s changed is that having the “best brand” will be much much harder in a world where any SaaS product can have a thousand competitors born the day after it launches — each of which will also be competing to be the best brand. Any modern-day category that has intense competition almost always becomes a story of price vs. brand. Smart buyers realize they can pin vendors against one another until the price spirals down.
What does the next decade look like?
We all have a tendency to look at the past decade of unicorns and because of some anchoring bias, believe that the next decade should look similar. But Alfred’s presence should necessarily imply that many of the past decade’s unicorns would not exist if they were started today. The next decade must look like a post-Alfred world, meaning that the best companies will be ones where value is accruing to some dimension that is not software creation. Maybe it’s some core service that the company does or some legal risk it takes on. In the same way that content creation went from something that large studios made to something that individuals do on their phones, the creation of usable and complex Software products will go from something that large teams do to small teams or individual creators.
AI eats Software
When we look back at this time we’ll see 2005 — 2023 as the age of Software eating the world, and we’ll look at 2023 — 2030 as the age of AI eating software and then after that the rest of the world. The near term might be one of the best times in history to be a bootstrapped small team building or be building a product where software is an enabler but not the core product.